Вы находитесь на странице: 1из 12

Labour Economics 27 (2014) 4455

Contents lists available at ScienceDirect

Labour Economics
journal homepage: www.elsevier.com/locate/labeco

Job polarization in aging economies


Eva Moreno-Galbis a,b,c,d,, Thepthida Sopraseuth e,f
a

GRANEM, University of Angers, France


CREST, France
c
IRES, Catholic University of Louvain, Belgium
d
GAINS-TEPP, University of Maine, France
e
THEMA, University of Cergy, France
f
Cepremap, France
b

H I G H L I G H T S
Goods and services are complementary for seniors and substitute for young people.
New technologies replace labor input in routine tasks.
Labor input reallocates towards the service sector in aging societies.

a r t i c l e

i n f o

Article history:
Received 27 August 2012
Received in revised form 14 November 2013
Accepted 8 December 2013
Available online 25 December 2013
JEL classication:
D12
J14
J21
O33
Keywords:
Job polarization
Elasticity of substitution
Aging
Technological change

a b s t r a c t
The progressive diffusion of ICT explains the raise in the number of highly paid jobs but has difculties in justifying that of low-paid jobs. Classifying occupations according to their median wage in 1993, we analyze their employment growth until 2010, which is highest both in the top and in the bottom of the distribution, and lowest in
the middle. Low-paid personnel services arise as the main factor responsible for the increase in the proportion of
employment at the bottom of the wage distribution. We argue that population aging can explain the increased
demand for personal services and thus the rise of employment in low-paid positions. Our argument goes as
follows: goods and personal services are complementary for seniors. The decrease in the relative price of
goods, induced by the progressive replacement of labor input in routine tasks by machines, is then associated
with an increased demand for personal services if the proportion of seniors is increasing. We thus complement
the existing literature on employment polarization by showing that demographic trends also play rst order role.
2013 Elsevier B.V. All rights reserved.

1. Introduction
This paper investigates the impact of aging on recent labor market
dynamics. Autor and Dorn (2013) show that employment changes in
the US since the mid-1990s are U-shaped, with relative employment
decreases in the middle of the wage distribution and relative gains at
the tails. At the bottom of the distribution, they nd that employment
gains are accounted for by growth in low-skilled service occupations;
We thank the two anonymous referees, Franois Langot, Bruno Decreuse, Ahmed
Tritah, Sylvie Blasco, and participants in the Louvain Economic Seminar and T2M
Conferences for their comments. All remaining errors are ours. Thepthida Sopraseuth
acknowledges the support of the Institut Universitaire de France.
Corresponding author at: GRANEM, University of Angers, France. Tel.: + 33 2 43
83 36 56.
E-mail addresses: Eva.morenogalbis@univ-angers.fr (E. Moreno-Galbis),
thepthida.sopraseuth@u-cergy.fr (T. Sopraseuth).
0927-5371/$ see front matter 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.labeco.2013.12.001

these jobs often involve assisting or caring for others (e.g. food service,
security guards, janitors, child care workers, hairdressers). Autor and
Dorn (2013) relate this job polarization to the combination of nonneutral technological progress and consumer preferences that favor variety over specialization. Non-neutral technological progress reduces
the cost of routine tasks since labor input is replaced by cheaper machines, but it has only a minor impact on the cost of work associated
with service occupations. This results in a decrease in the relative
price of goods with respect to services. If goods and services are complements the demand for service occupation outputs will increase,
resulting in an increase in both employment and wages in service occupations. Autor and Dorn (2013) are silent about the potential reasons
behind this complementarity in consumers' preferences. They also fail
to explain the absence of labor market polarization before the mid1990s even though American technological diffusion started in the
1980s. We believe that population aging explains both of these

E. Moreno-Galbis, T. Sopraseuth / Labour Economics 27 (2014) 4455

phenomena. Because goods and personal services are complements for


seniors, population agingin conjunction with technological progress
increases demand for these services. Further, labor market polarization
appeared when population aging started to become an issue for most
western economies.
We propose a simple theoretical framework following Baumol
(1967) to test our hypothesis. Our model provides an equation to estimate the elasticity of substitution between goods and personal services
by age group. It further allows us to analyze the combined impact of
population aging and technological progress on demand for labor
input in the personal service sector.
Our contribution is twofold. First, we estimate age-specic elasticities of substitution between goods and personal services. If the elasticity
of substitution is constant, then aging is sufcient to produce job polarization at the bottom of the wage distribution and allows for increased
complementarity (with an upper bound). These elasticities suggest
that goods and personal services are indeed complements for older
households while being substitutes for young ones.
Second, in the context of technological change between 1993 and
2010, we estimate the contribution of population aging to the increased
concentration of jobs at the bottom of the wage distribution. We provide evidence of the gradual polarization1 of the labor market and nd
that demand for service occupation outputs drives employment growth
at the lower tail of the distribution while driving up wages for this type
of job.
To our knowledge, this is the rst paper that presents evidence of job
polarization in French data with a stress on employment growth in the
service occupation sector. Dealing with the relationship between polarization and the wage distribution is left for future research. Our objective here is to analyze the combined contribution of aging and
technological change to the job polarization process, in terms of employment changes.
The paper is organized as follows. Section 2 is a literature review.
Section 3 sketches the theoretical framework. Section 4 empirically validates our assumption concerning the divergent elasticity of substitution between goods and personal services for seniors and young
workers. Section 5 provides evidence of job polarization in the French
labor market. It shows that employment demand for service occupations has increased due to the combined effect of technological change
and aging. Section 6 concludes.
2. Literature review
The literature is far from unanimous concerning the reasons behind
the increased demand for labor input in non-routine manual jobs. Initially, most of the literature simply focused on the progression of wage
inequalities. Using US data, Lemieux (2006) and Autor et al. (2008)
found that wage inequalities between the ninth and the fth decile
had increased more than wage inequalities between the ninth and the
rst decile, suggesting a rising demand for jobs at the two extremes of
the distribution or, at least, a decreasing demand for jobs in the middle
of the distribution.
Autor et al. (2006) for the US, Spitz-Oener (2006) for Germany,
Maurin and Thesmar (2004) for France, and Goos and Manning
(2007) for the UK nd that labor input in routine positions (those in
the middle of the wage distribution) has been gradually replaced by
cheaper and more productive machines. This automation improved
the productivity of labor input in abstract tasks (those at the top of the
wage distribution) by facilitating access to information. On the other

1
By denition, a polarization process implies an increasing concentration in both extremes of a given distribution. In this paper, we will often refer to polarization at the bottom of the distribution or polarization at the top of the distribution, meaning that the focus
of the analysis is on the bottom or the top tail of the wage distribution, even if we are
aware that polarization must concern both extremes of the distribution.

45

hand, this new technology created a mass of jobless medium-skilled


workers. Part of this mass may remain unemployed,2 whereas another
part may be reallocated towards more labor-demanding sectors. Obviously, in spite of the increased demand for qualied labor in abstract positions, there is a skill mismatch since medium-skilled workers do not
have the required qualications to apply for these types of nonroutine jobs. They will thus have to switch to manual, non-routine positions (see Goos and Manning (2007)). The productivity for these positions has remained unaffected. These types of position are difcult to
automate or outsource since they require interpersonal and environmental adaptability as well as direct physical proximity. But why should
the demand for these jobs rise in such a way as to absorb part of the
workers that have been replaced by machines? Our paper points to population aging as an answer to this question.
Using British data, Manning (2004) has already observed that the
employment of low-skilled workers is increasingly dependent on their
physical proximity to high-skilled workers, as low-skilled work is increasingly concentrated in the non-traded sector. According to him, if
there are relatively more skilled workers in a city, we could expect the
demand for unskilled labor in the non-traded sector to rise. This
would lead to fewer unskilled workers in the traded sector and an overall increase in demand forand wages ofthe unskilled. Within a
delimited geographical space, a kind of complementarity between unskilled and skilled workers seems to arise. The biased nature of technological progress must have directly promoted skilled labor and
indirectly (via the complementary relationship) unskilled labor.
Based on US data, Mazzolari and Ragusa (2013) provide an explanation similar to Manning (2004)'s of the increasing demand for low-paid
services. Their idea is that low-skilled workers are employed in nontradeable, time-intensive services that are substitutes for home production activities. Over the past decades, wage gains at the top of the wage
distribution have increased the opportunity cost for high-skilled
workers in spending time on domestic activities. Rather than
implementing this domestic production themselves, high-skilled
workers now prefer to buy these home services. So the increase in demand for low-paid services at the bottom of the wage distribution results from the combination of a substitution effect (the opportunity
cost of one hour of home production is now higher for high-skilled
workers) and an income effect (because high-skilled workers are
wealthier, they can buy more services). In sum, Mazzolari and Ragusa
(2013) and Manning (2004) conclude that rising wage inequality between highly paid skilled workers and unskilled workers will induce
the formers to demand more low-skilled services so as to free up
more of their time for market work.
Based on 16 European countries, Goos et al. (2009) also analyze the
importance of income inequality effects. They conclude that the relative
growth in low-paid service occupations cannot be explained by the increase in real income and non-homothetic preferences. The higher income elasticities they estimate are actually associated with the top of
the wage distribution (more precisely, with the three high-paid industries: (i) nancial intermediation, (ii) real estate, renting and business
activity, and (iii) transport, storage and communication). Goos et al.
(2009) nd that the automation hypothesis remains the most important factor behind the observed shift in the employment structure. Little
support is found for the hypothesis that changes in product demand are
driven by income inequality effects.3 Further they nd a relatively small,

2
Cheron et al. (2011) show that the diffusion of new technologies has fostered a gradual increase in the relative unemployment rate of medium-skilled workers.
3
Rather than focusing on income inequality effects, Clark (1957) considers a uniform
income effect. He argues that the income elasticity of the demand for personal services
is greater than one, in which case, a general rise in income will tend to shift employment
towards service-intensive occupations. This pure income effect is not considered by more
recent papers, which focus rather on the impact that the recent increase in income inequality has had on job polarization.

46

E. Moreno-Galbis, T. Sopraseuth / Labour Economics 27 (2014) 4455

negative effect of globalization on demand for offshorable jobs; the real


driving force of this decreased demand is technological change.
Autor and Dorn (2013) conrm these ndings for US data. More precisely they consider three potential explanations for the rise in service
occupation employment: 1) offshoring, which displaces low-skilled
workers into non-offshorable service occupations; 2) rising income at
the top of the wage distribution, which stimulates demand for personal
services by wealthy households; and 3) rising returns to education,
leading college-educated workers to increase the labor supply and substitute market for home-based production of household services. They
nd that these mechanisms are not empirically important drivers of polarization. They further stress that increased automation fosters a reduction in the price of goods relative to the price of personal services. This
leads to an increase in demand for personal services only if goods and
personal services are complements. Biased technological progress is
necessary, but not sufcient, to explain job polarization; on its own, it
can only justify polarization at the top of the distribution as it increases
the productivity of abstract tasks. To explain polarization at the bottom
of the distribution, where the productivity of labor remains unaffected
by ICT diffusion, we also require goods and personal services to be, at
least, weakly complementary. In this way, biased technological progress
fosters a reduction in the relative price of goods.
To summarize, in order to understand the concentration of jobs at
both extremes of the wage distribution, we need to combine biased
technological change with a complementary relationship between
goods and personal services, as in Autor and Dorn (2013). Our originality lies in our stress on aging, which allows us to understand the time lag
between technological diffusion and polarization. Because the diffusion
of skill-biased technologies started in the 1980s, whereas the polarization process is a more recent phenomenon (beginning in the second
half of the 1990s), we argue that the aggregate elasticity of substitution
between goods and personal services has decreased with respect to past
decades, when we observed biased technological change but no polarization. Our paper therefore adopts a novel perspective by pointing to
the combined effects of population aging and technological progress
as the main drivers of the increase in demand for personal services, controlling for the potential effect produced by the increase in income
inequality.

utility function that the representative household maximizes subject


to the aggregate budget constraint. More precisely,

 h
i

1=
maxqgt ;qst U qgt ; qst 1as qgt as qst
subject to
pgt qgt pst qst yt ;

3.1. The household decision


3.1.1. Consumption decision
For simplicity, we assume a large representative household composed of individuals who might be retired, unemployed, or employed
in either the goods or personal services sector. Each individual has a different elasticity of substitution between goods and personal services.
However, the aggregation of all individual preferences4 yields a single
4
It is beyond the scope of this paper to deal with aggregation issues since the unique
goal of the theoretical model is to provide a framework giving a rationale to the results obtained in the following sections.

where qgt denotes the quantity of goods, qst the quantity of personal services, and pgt and pst, their prices. as N 0 is the weight of personal services in the consumption basket. The household's income is yt and
comes from wages earned by employed individuals and the retirement
pensions of seniors. The household's aggregate elasticity of substitution
1
between goods and personal services is given by 1
. Replacing the
rst order conditions in the budget constraint yields the following demand for personal services:
1
1

qst as

1
1

pst
yt ;
P t

h
i
. Taking logwith the price index denoted P t 1as pgt as pst
arithms on both sides of Eq. (3) gives the following log-linear demand
for services:
1
1

1
1

log qst log yt log pst 1 log P t log as :

A log-linear approximation of the aggregate price index is:


log P t log pgt 1 log pst ;

where is the share of income spent on the routine good sector. In other
words, approximates the size of the goods sector. Substituting Eq. (5)
in Eq. (4) yields:
log qst log yt log pst 1 log pgt 11 log pst log as :

3. A simple theoretical model


Suppose an economy can be broadly divided into two sectors: the
routine goods sector, employing labor and capital to produce goods,
and the low-skilled, personal service sector, employing labor to produce
value. Because routine tasks in the goods sector are easily automated,
they can be implemented by capital or labor. Both factors are assumed
to be perfectly substitutable in the production process of routine
goods. Moreover, technological progress positively affects the relative
productivity of capital. Finally, note that technological progress does
not affect the marginal productivity of labor in the personal service sector. It will, however, decrease the relative productivity of labor with respect to capital in the goods sector.

b1

3.1.2. Labor supply decision


Young members of the household inelastically supply one unit of
labor, though they may be employed or unemployed. Participation issues are not considered, so we assume that all young individuals participate in the labor market.5 Moreover, due to perfect labor mobility
between sectors, wages in the two sectors are equal. Individuals are
therefore indifferent between working in one sector or the other.
3.2. Production decisions
We also suppose a perfectly competitive environment for the goods
and services markets but an imperfect competitive environment for the
labor market. Firms set prices for goods and services equal to their respective marginal costs. These marginal costs are closely linked to
wages paid by the rm. Wages are assumed to be the result of a
bargaining process. In the routine goods sector, wages equal the marginal productivity of labor. In the low-qualied personal services sector,

5
Being unemployed or not participating in the labor market is equivalent in this context. This is because there are no unemployment benets. Unemployed people simply
generate domestic production. One might object that the participation margin has been
shown to play a role in accounting for cyclical unemployment. However, in this paper,
we focus on long run labor market adjustments. In France, during the period considered
in our econometric exercise, the participation rate of young workers has remained fairly
constant at 41.2% with a standard deviation of only 1.2 points (Source: INSEE).

E. Moreno-Galbis, T. Sopraseuth / Labour Economics 27 (2014) 4455

we assume that the rm pays the worker a wage set to his outside
opportunity of employment.6
In the routine goods sector, technological progress increases the
relative productivity of capital, which boosts aggregate production. As
capital and labor are substitutable in the goods sector, the increasing relative productivity of capital encourages rms to increase their demand
for capital with respect to labor. The subsequent increases in production
are due to this increased use of capital. The production function in the
routine good sector is:



1=
gt
qgt aKt K t aLt Lgt
with aKt K e and aLt aKt 1

where Kt and Lgt are capital and labor, while the growth rate of technological progress is given by g. An increase in g improves the relative productivity of capital, which should lead to a rise in the demand for this
production factor. As in Autor and Dorn (2013), we normalize the
good price to one: pgt = 1.7
While technological progress raises productivity in the goods sector,
productivity in the low-skilled services sector, which employs only
labor, remains unaffected. This is consistent with the view that personal
services require interpersonal and environmental adaptability as well as
direct physical proximity. The productivity of labor in this sector does
not depend on capital intensity. We normalize this productivity to
unity as in Autor and Dorn (2013). The production function then takes
on a very simple form:
qst Lst :

Because wages in the services sector equal the value of the outside
opportunity of employment, the marginal cost in the services sector
equals: b egt, where b can be interpreted as the value of leisure or
home production. The value of the outside opportunity of employment
increases at the same pace as the aggregate productivity of the economy. The price set by this sector is then given by pst = b egt. Replacing
pgt = 1 and pst = b egt in Eq. (6), and applying Eq. (8), yields:
log Lst C log yt 1gtgt;

where C is a constant. Since the total labor supply is normalized to 1 (i.e.


Lst + Lgt = 1), Lst is the labor share in the service occupation sector. In
Eq. (9), income positively inuences the employment share in personal
services. The impact of technological progress depends on the degree of
substitutability between goods and personal services. On the one hand,
there is a direct negative impact equal to gt of growth on employment
in personal services. It results from the increase in personal service
wages associated with the rise in g. On the other hand, the impact of
(( 1))gt depends on the value of . If goods and personal services
are complements (0 b b 1), the term (( 1))gt becomes positive, meaning that technological progress increases the demand for personal services and, hence, employment in this sector. The importance of
this effect is scaled by ; the larger the , the more signicant the effect.
If goods and personal services are substitutes then N 1; then the term
6
With Nash bargaining on wages, our model is equivalent to a setting in which workers
have different bargaining powers across sectors. In the routine goods sector, workers are
assumed to have much more bargaining power than in the personal services sector. This
is because wages equal the marginal productivity of labor in the goods sector while, in
the personal services sector, rms pay a wage equal to the workers' outside opportunity
of employment. This is consistent with the empirical evidence. Hirsch and Berger (1984)
nd evidence of larger union density in more capital-intensive rms. Further, using the
2013 release of the Data Base on Institutional Characteristics of Trade Unions, Wage Setting, State Intervention and Social Pacts (ICTWSS), we nd that the average union density
is 40% in manufacturing industries versus 1% in hotels and 4% in household services. This
average was calculated using the latest observations for Australia, Austria, Belgium, France,
Germany, Spain, Ireland, Italy, Switzerland, and the US. This suggests that workers in the
goods-producing sector have much more bargaining power than their counterparts in
the personal services sector.
7
Since we have a real model, only relative prices matter.

47

(( 1))gt is negative. This implies that technological progress reduces employment in personal services. This reduction is amplied as
increases.
Autor and Dorn (2013)'s routinization hypothesis, according to
which the larger the share of routine jobs the more marked the job polarization process, contributes to job polarization at the bottom of the
job quality distribution when goods and personal services are complements. The question that arises then is: when are goods and personal
services complements rather than substitutes? The age composition of
the population determines the aggregate elasticity of substitution between goods and personal services (see Appendix A). Combining population aging with the progressive replacement of labor input in routine
tasks by relatively more productive machines explains the increase in
the proportion of employment at the bottom of the wage distribution.
Our model has a number of testable implications. First, a reduced
form of Eq. (4) is estimated so as to compare the value of the elasticity
of substitution between goods and personal services for households belonging to different age groups. This is done in Section 4. Second, a reduced form of Eq. (9) will allow us to estimate the combined impact
of aging and automation on the demand for low-skilled personal services (Section 5). While the share of seniors does not explicitly appear
in Eq. (9), we can consider this share as a proxy of the degree of complementarity or substitutability between goods and personal services (see
Appendix A).
4. Estimating the elasticity of substitution between goods and
personal services
Our main claim is that, combined with technological progress, population aging induces an increase in demand in the low-skilled personal
service sector (i.e. job polarization at the bottom of the wage distribution) due to the complementary relationship between goods and personal services. More precisely, we argue that for seniors good and
services are more complementary than for young workers. Since recent
technological changes have reduced the price of goods, demand for personal services will tend to increase if goods and personal services are
complementary. Conversely, we expect a reduction in the demand for
services if they are substitutes with respect to goods.
4.1. The 2006 Household Budget Survey
The aim of the Household Budget Survey (HBS) is to compile data on
the expenditure and resources of French households. The study of expenditure is the central purpose of the survey. All household expenditure is recorded. The nature of these expenses is broken down into a
classication of about 900 budgetary items compatible with the classication used in the national accounts. All expenses are covered, including
those not associated with the consumption of goods and personal services.8 In addition to expenses, the survey collects information about
non-monetary consumption (e.g. food produced for own consumption
and imputed rentals). Essentially, the family household budget survey
collects monetary data, leaving the more specialized surveys on each
item of consumption (e.g. transportation, housing, leisure, or holidays)
to use a more qualitative approach to household behavior. Nevertheless,
in order to illustrate these monetary data, some complementary questions are asked about the nancial situation as perceived by the
household.
The statistical unit is the household and all ordinary French households are covered.9 The survey began in 1979 and takes place approximately every 5 years. Since 2001, expenditure has been classied
according to the European classication, which differs considerably
8
The precise expenses are: taxes and contributions, insurance premiums, major home
renovation expenditure, inter-household transfers, purchase of second-hand goods and
loan repayments.
9
This includes the overseas departments.

48

E. Moreno-Galbis, T. Sopraseuth / Labour Economics 27 (2014) 4455

from the French classication, which was used until 1995. For this reason, the results of the 2001 and 2006 editions of the survey cannot yet
be compared directly with those from previous years. Due to the absence of a panel and since expenditures on personal services are only
detailed in the 2006 survey, we employ it in our study. The 2006 (random, self-weighted) sample consists of about 20,000 dwellings in mainland France and 5000 in the overseas departments. The number of
households surveyed is about 10,240 for mainland France and 3134
for the overseas departments.

and 114 young households. For households (young and old) whose
cleaning expenditure equals zero we set the hourly price equal to the
2006 minimum wage of 6.9285. This value seems quite reasonable
given that the median of the sample equals 6.11 and the mean 9.46.
4.3. Estimates of the elasticity of substitution
We now have all the required information to estimate the reduced
form of:
log qst log yt log pst 1 log P t log as :

4.2. Computing the price and expenditures for personal services in


the survey

Due to data availability problems, we assume that the general con#


"
1

We consider senior households to be those composed of one or two


people older than 60. We eliminate all senior households where a younger individual resides. This is because the presence of the younger individual may skew the older household's demand as he could help out
with services that might otherwise be purchased. We have more than
2540 senior households. We dene a young household as one composed of one or two 3045 years old who do not have children. We
have more than 1800 young households.
The survey provides information on the annual expenditures of hiring someone for personal services.10 Most households, particularly
young ones, declare zero expenditure on the majority of these services.
House cleaning is the expenditure for which the number of households
declaring a strictly positive quantity is the highest; 472 senior households and 120 young households declare strictly positive house cleaning
expenditure.11 Thus, we will focus on this type of personal service. Additional robustness tests where we consider all expenditures on personal services are presented in Appendix B.1 and conrm the results
presented in this section.
The rst problem we must deal with is the lack of information
concerning the number of hours worked by the personal service employee. There is only one variable in which households declare the number of hours hired over the previous two months. However, there are a
great number of missing observations in this variable. We solve both
problems sequentially. First, since we deal with yearly expenditures,
in order to compute the hourly cost we assume that the number of annual hours worked by a personal service employee is six times the number of hours worked over the two month period. Next, to deal with the
problem of missing data, we compute the average number of hours
hired, grouping households with comparable yearly expenditures together. These averages are then imputed to those households with
missing data that have similar personal expenditure levels. Clearly,
this process assumes that households with comparable house cleaning
expenditures hire a similar number of hours.
Finally, to compute the hourly price paid by households for personal
services, we divide yearly expenditure by yearly hours (declared or imputed). Obviously, since yearly hours are computed by extrapolating the
number of hours declared for the two months preceding the survey to
the whole year, the hourly price is, in many cases, unreasonable. We
drop these unreasonable observations.12 The nal sample of households
declaring a strictly positive house cleaning expenditure (and thus paying a strictly positive hourly price) is reduced to 425 senior households
10
Personal services considered in the survey are house cleaning, child care, shopping,
domestic health care, and garden care.
11
For child care there are no senior households declaring a strictly positive expenditure
and there are 32 young households declaring strictly positive expenditures. For shopping
there are 24 senior households and zero young households. For domestic health care 8 senior households and 1 young household declare a strictly positive expenditure. For gardening there are 71 senior households and 7 young households. For other services there
are 24 senior households and 1 young household.
12
We believe that an hourly price above 150 is too high to be realistic and an hourly
price below 1 is far too low, even for households for which the cost of these personal services may be subsidized. There is always a marginal cost borne by the household. This happens with most medicines, which are greatly, but not fully, subsidized.

1
1

1
1

sumption price index P t 1as pgt as pst

is constant across

13

individuals. This term together with the weight as is then included in


the constant. Our estimation reduces to:
log qsi C log yi log psi ;

10

where i represents the individual household, stands for the elasticity of


substitution, C for the constant, y for income, ps for the hourly price of
house cleaning services and for the error term. Income14 is computed
from the database and varies across households. The hourly price of personal services also varies among households and results from the computations explained above. For each type of household, we implement
weighted OLS estimation with and without geographical dummies so
as to test the robustness of the results. The coefcients are stable,
which makes us quite condent about our results. Moreover, by
implementing weighted regressions, we eliminate the potential problem associated with an over-representation of certain types of households. The data do not allow us to determine the precise location of
the household. They do however provide the geographical area of the
household's main residence (Mediterranean zone, Center-east, Southwest, West, East, North, Paris metropolitan region) and the type of city
of residence (rural area, city of less than 20,000 inhabitants, city having
between 20,000 and 100,000 inhabitants, city of more than 100,000 inhabitants, city of Paris). We interact both variables so as to obtain more
precise geographical dummies (40 dummies).
Table 1 summarizes our results. Columns 1 and 2 provide the estimation of the elasticity of substitution (coefcient on ps) for senior households when we only consider those households declaring positive
house cleaning expenditures. The corresponding elasticity for young
households is in columns 5 and 6. The size of the sample is quite small,
particularly for young households where there are only 113 observations. This makes us very cautious when interpreting the coefcients.15
Columns 3 and 4 include results for all senior households. The corresponding results for young households are contained in columns 7 and
8. Here, the size of the sample increases to 2362 observations for seniors
and 1802 observations for young workers. However, it must be noted
that only 17% of the observations for seniors and 6% of the observations
for young households are strictly positive. Our sample is thus leftcensored.16 The estimated elasticity is signicant and below unity for
13
Due to the great amount of missing information on the large number of expenditures
made by the household, trying to estimate the particular price index per household would
introduce a major source of measurement error in our model.
14
Income refers to permanent income and results from the sum of the retirement pension (if any), salary or income from self-employment.
15
Estimated elasticities are below one. However, these elasticities remain above the
elasticities estimated for senior households over the reduced sample (columns 12). This
suggests that goods and personal services are more complementary for senior households
than for young households.
16
In order to deal with left-censored observations, we have also implemented a Tobit estimation which is available from the authors' web appendix (http://thepthida.sopraseuth.
free.fr, https://sites.google.com/site/morenogalbiseva). Estimated coefcients correspond
to the average marginal effects (they can no longer be interpreted as elasticities) and they
appear to be lower for seniors than for young households.

E. Moreno-Galbis, T. Sopraseuth / Labour Economics 27 (2014) 4455

49

Table 1
Elasticity of substitution between goods and personal services (house cleaning). Budget Household Survey 2006.
Dependent variable: demand for house cleaning services

Log y
Log ps
Geograph. dummies
Observations
R-squared
VIF

Senior households with


positive expenditure

All senior households

Young households with


positive expenditure

All young households

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

0.0266***
(0.000562)
0.103***
(0.000475)
Yes
400
0.105
12.04

0.0345***
(0.000566)
0.0900***
(0.000469)
No
400
0.031
1.07

0.135***
(0.000870)
0.904***
(0.00163)
Yes
2362
0.063
16.93

0.146***
(0.000867)
0.900***
(0.00162)
No
2362
0.045
1.01

0.212***
(0.00163)
0.267***
(0.00115)
Yes
113
0.325
7.43

0.177***
(0.00152)
0.214***
(0.00117)
No
113
0.19
1.00

0.326***
(0.000821)
1.232***
(0.00236)
Yes
1802
0.120
13.77

0.322***
(0.000806)
1.243***
(0.00237)
No
1802
0.104
1.000

Standard errors in parentheses: p b 0.01.

seniors while it is signicant and above unity for young households.


That is, goods and personal services are complements for seniors and
substitutes for young households.
What is remarkable in Table 1 is the stability of the coefcients independent of the introduction of geographical dummies. The inclusion of
geographical dummies induces a multicollinearity problem (the variance
ination factor, VIF,17 is above 5) that disappears when we eliminate geographical dummies in columns 2, 4, 6, and 8. However, the value of the
coefcients associated with log ps and log y in columns 1 and 2, 3 and
4, 5 and 6, and 7 and 8, remains quite close and the degree of signicance
unchanged. The multicollinearity concern induced by geographical
dummies does not then seem like a major problem in the estimation of
the elasticity of substitution. Appendix B.2 proposes a similar exercise
on a pooled sample of young and senior households. Results remain robust. Goods and personal services are complements for seniors.
5. Job polarization and aging in French microdata
The polarization of the labor market has been largely documented
for some OECD countries.18 In this section we show that this polarization process also applies to the French case. Next, we try to gain insights
into the type of jobs concentrated at the bottom of the wage distribution. Personal services arise as the main determinant of employment
growth at the lower tail of the distribution. Finally, after investigating
a possible gender composition effect behind job polarization at the bottom of the distribution, we perform an econometric analysis of the role
of population aging, in the context of technological change, on this job
polarization process. A reduced form of Eq. (9) is then estimated.
5.1. The French Labor Force Survey
The French Labor Force Survey (LFS) was launched in 1950 and
established as an annual survey in 1982. Redesigned in 2003, it is now
a continuous survey providing quarterly data. Participation is compulsory and it covers private households in mainland France. All individuals
in the household older than 15 are surveyed.
The quarterly sample is divided into 13 weeks. From a theoretical
point of view, the sampling method consists of a stratication of mainland France into 189 strata (21 French regions 9 types of urban unit)
and a rst stage sampling of areas in each stratum (with different

17
The VIF quanties the severity of multicollinearity in an OLS regression analysis. It provides an index that measures how much the variance of an estimated regression coefcient is increased because of collinearity. A VIF below 5 implies no multicollinearity
problem.
18
See Goos and Manning (2007) for the UK, Autor et al. (2006) for the US, Spitz-Oener
(2006) for Germany or Goos et al. (2009) for 16 European countries.

probabilities, average sampling rate = 1/600). Areas contain about 20


dwellings and among them only primary residences are surveyed.
Each area is surveyed over 6 consecutive quarters. Every quarter, the
sample contains 6 sub-samples: 1/6 of the sample is surveyed for the
rst time, 1/6 is surveyed for the second time, and 1/6 is surveyed for
the 6th (and last) time. When it was run as an annual survey, every
year a third of the sample was renewed meaning that each individual
was interviewed only 3 times. The collection method has always been
a face-to-face interview.19
Topics covered by the LFS concern employment, unemployment, underemployment, hours of work, wages, duration of employment and
unemployment (length of service), discouraged workers, industry, occupation, status in employment, education/qualication, and other
jobs. The French LFS provides the occupation for each employed individual among a list of 350 possible occupations such as gardener, messenger, clerks in banking activities and nancial managers.20
Some occupations are characterized by a very general denition and a
large number of employed individuals. In this case, we disaggregate
these jobs by industry.21 For instance, secretaries are divided into secretaries in the food industry, in the car industry, etc. This leaves us with
452 occupations each year.
5.2. Evidence of job polarization in France
The French LFS allows us to have a look at trends in the quality of
jobs (i.e. employment structure). As in Goos and Manning (2007), quality is proxied by the median hourly wage for each job at the beginning of
the period.
We rst relate the job quality in 1993 to the average subsequent
change in log employment from 1993 through 2010.22 Fig. 1 presents
the average change in log employment in the 19932010 period
(Y-axis) for each job quality, proxied by its median wage (X-axis).23 Circle sizes denote the employment level in 1993. We also present a quadratic t of average employment growth by job quality. Panel (i) of

19
Since 2003, a telephone interview has been employed for intermediate surveys (2nd
to 5th).
20
With this range of possible jobs, the wage distribution captures both low and high paid
jobs. However, the number of observations in top jobs is particularly low in all surveys. The
sample does not capture the increase in wage inequality due to the top 1% of the wage distribution (Landais, 2008).
21
Appendix C provides a description of the occupations.
22
Unfortunately, we could not use the LFS prior to 1993 because of a drastic change in
industry classication. Thus it was not possible to obtain consistent industry codes over
time.
23
For the sake of clarity, all graphs in this section have been rescaled by removing outliers from the chart. The outliers are not excluded in computing the quadratic regression
curves. The web appendix (http://thepthida.sopraseuth.free.fr, https://sites.google.com/
site/morenogalbiseva) presents graphs including outliers.

50

E. Moreno-Galbis, T. Sopraseuth / Labour Economics 27 (2014) 4455

(ii ): No personal services

.5

.4

.2

.2

.4

.5

(i ): Benchmark

7.5

8.5

9.5

10

7.5

Log median wage in 1993


Average change in log employment

8.5

9.5

10

Log median wage in 1993


Fitted values

Average change in log employment

Fitted values

.1

.1

.2

.3

(iii ): Wages

7.5

8.5

9.5

10

Log median wage in 1993


Average change in log median wage

Fitted values

Fig. 1. Average employment and wage growth by job median wage (19932007). X-axis: Log-median wage in 1993. The size of the circles corresponds to the employment level in 1993.
Y-axis: Average growth rate of employment between 1993 and 2010.

Fig. 1 suggests a U-shaped relationship between employment growth


and job quality.24 The quadratic t can be represented as follows:
2

d logemployment logw1993 logw1993 C


2
d logemployment 0:6384738
logw1993  0:0373618  logw1993  2:734133 

 


0:0000521
0:0040663
0:0009208

11
where standard errors are presented in parentheses.25 While the coefcient of the linear term, log w1993, is negative and signicant, the coefcient associated with the quadratic term, (log w1993)2, is positive and
signicant, conrming a U-shaped employment structure prevailed between 1993 and 2010. Over the past decades, jobs in the middle of the
wage distribution have been characterized by low employment growth
while the rise in employment has been greater for jobs at the top and
bottom of the wage distribution.
At the lower end of the job quality distribution, two jobs are characterized by a high employment level in 1993 and strong subsequent employment growth. These large circles correspond to occupations related
to child care and house cleaning or, more generally, personal service occupations. This is clearly observed in Panel (ii) of Fig. 1 which reveals
24
We choose to consider the change in log employment over the 19932010 period, as
in Goos and Manning (2007). Autor and Dorn (2013) examine the change in employment
share. We observe that the relationship between job quality and change in employment
share is still U-shaped, thereby suggesting job polarization.
25
(): signicant at 1%. (): signicant at 5%. (): signicant at 10%.

that, if personal service occupations are excluded from the analysis,


the U-shape curve disappears.26
Panel (iii) shows that the increased demand for jobs at the bottom of
the wage distribution, has caused wages there to increase more than for
other jobs.27 The three gures together send a clear message: between
1993 and 2010, positions at the bottom and at the top of the wage
distribution have beneted from larger increases in employment than
positions in the middle. This is particularly true for personal services,
which have experienced the biggest increases in both employment
and wages.
Finally, various factors could potentially cause this concentration of
jobs at the bottom of the wage distribution. While our paper focuses
on the combined effect of technological progress and aging, feminization of the labor force could also be a source of this job polarization,
with women accounting for the growth in relatively low-paid occupations. But, as Fig. 2 shows, one observes similar patterns for male and female employment considered separately in the 19932010 period. The
26

Actually, the quadratic t of the bottom part of the job distribution now equals:
2

d log employment 0:2334882 log w1993 0:0135525 log w1993 1:005486


0:0047966
0:0002856
0:0201244
suggesting an inverted U-shape progression in this part of the distribution.
27
The quadratic t of the wage distribution equals:
2

d log wage 0:0450597 log w1993 0:0018549 log w1993 0:2754366


0:0001888
0:0000107
0:0008338 :

E. Moreno-Galbis, T. Sopraseuth / Labour Economics 27 (2014) 4455

(ii ): Benchmark (men)

.5

.4

.2

.2

.4

.5

(i ): Benchmark (women)

51

10

Log median wage in 1993


Average change in log employment

10

Log median wage in 1993


Fitted values

Average change in log employment

Fitted values

Fig. 2. Average employment growth for men and women by job median wage (19932007). X-axis: Log-median wage in 1993. The size of the circles corresponds to the employment level
in 1993. Y-axis: Average growth rate of employment between 1993 and 2010.

results do not change with respect to the pooled sample. The quadratic
t conrms this result for both men and women:
2

d log female employment 1:048863


log w1993 0:0614585
log w1993
 4:472667  
 
 
0:0017161
0:0001
0:0073536
2

d log male employment 0:2995561


log w1993 0:0181389
log w1993  1:241833 
 
 

0:0001138
0:0094687
:
0:0020772

In sum, three conclusions can be drawn from Figs. 1 and 2 that are
consistent with the literature presented in Section 2. First, over the
past decades the French labor market has become more polarized. Second, as shown in Panel (ii) of Fig. 1, the main driver of job polarization at
the bottom of the wage distribution is personal services.28 Finally, one of
the main explanations for the increase in the share of personal services
is the feminization of the labor force. However, empirical evidence contradicts this hypothesis since, regardless of whether we consider the
male or female labor force, the polarization process arises. Apart from
the integration of women into the labor market, one of the major demographic changes that are likely to have inuenced the labor market
structure over the past years is population aging. We would thus like
to evaluate the role of population aging over the past 18 years on job
polarization at the bottom of the wage distribution.
5.3. Change in service employment share and aging
Our estimation is based on the time-derivative of Eq. (9), that is:
:

Lst yt 1gg:

12

We will proxy the aggregate elasticity of substitution between goods


and personal services in a particular geographical unit by the share of
seniors (see Appendix A). As in Autor and Dorn (2013), we proxy technological change by the share of routine employment in the geographical unit.29 We identify routine occupations in the list of our 452 jobs.
Unlike other papers on the subject, we do not have a Dictionary of Occupational Titles allowing us to determine the task composition of each
occupation. We are thus obliged to classify jobs as either routine or
28

These results are conrmed by Table 5 in Appendix D.


Autor and Dorn (2013) nd that the share of routine employment in a commuting
zone at the beginning of the decade is highly predictive of computer adoption during that
decade.
29

non routine (see Appendix E for details). For our benchmark regression,
we adopt a tight denition of routine (so as to be sure that all jobs considered mainly include routine tasks). Control variables will not be introduced in variation so as to avoid causality problems. The value of
these variables will be xed at their initial level.
Our empirical investigation relies on a panel of small geographical
units: the French departments. We consider 96 departments included
in 21 regions. Our sample spans 19932010, which yields panel data
of more than 1700 observations. Looking at the department level rather
than the regional level allows us to consider a larger sample. A further
benet is that the empirical exercise will measure whether the relationship is pervasive over smaller geographical units. Moreover, this approach is consistent with the commuting zone approach of Autor and
Dorn (2013).
In each department, for each year, we compute the share of service
employment in total employment. We then compute the yearly change
in this share. Finally, we employ these values to compute the change in
the share of services employment in total employment for the periods
19932001 and 20022010. Since we have data on 18 years, we consider intervals of 9 years because we believe this period is long enough to
allow for an adjustment of the employment structure following a modication in the demand due to a demographic change. For each department, each of the two intervals constitutes our dependent variable.
With 96 departments and two changes in the service employment
share (19932001 and 20022010) this yields panel data of 192
observations.
Our empirical exercise is close to that of Autor and Dorn (2013) and
extends their analysis to French data. We would like to draw attention
to the demand shift associated with the combined effect of aging and
technological change as the main determinant of job polarization at
the bottom of the distribution. The estimated equation is:
service

Ni;t
Ni;t

!
oldi;t1  routinei;t1 X i;t1 ei;t ;

13



denotes the change in service employment share
where d NN


N
N
in department i and period t (t 1). Appendix E proN N
service
i;t
i;t

service
i;t

service
i;t1

i;t

i;t1

vides the list of service occupations. We adjust standard errors for clustering of observations at the department level.
Our paper explores the view that population aging may lead to a rise
in the share of service employment in a context of technological

52

E. Moreno-Galbis, T. Sopraseuth / Labour Economics 27 (2014) 4455

Table 2
Change in service employment share: benchmark estimation. Department level (19932010).
Dependent variable: change in service employment share
(1)
Oldt 1

(2)

(3)

(4)

(5)

(6)

0.126*
(0.0681)

0.128*
(0.0683)

0.151**
(0.0644)
0.00353
(0.0519)

0.00817
(0.0114)

0.0278***
(0.00909)
0.0141
(0.0119)

0.149**
(0.0651)
0.00315
(0.0517)
0.00514
(0.0105)
0.0282***
(0.00920)
0.0142
(0.0122)

Yes
Yes
0.0319**
(0.0126)
192
0.558
2.70

Yes
Yes
0.0253**
(0.0119)
192
0.581
4.12

Yes
Yes
0.0246**
(0.0121)
192
0.582
4.38

0.0765
(0.0634)

Routinet 1

0.0278*
(0.0157)

Oldt 1 Routinet 1
Unemploymentt 1
Femalet 1
Manufacturet 1
Managerst 1
Fixed effects:
Year
Department
Prop. min. wage
Observations
R-squared
VIF

Yes
Yes
0.0314**
(0.0129)
192
0.544
5.16

Yes
Yes
0.0333**
(0.0127)
192
0.555
2.34

Yes
Yes
0.0326**
(0.0126)
192
0.556
2.38

Standard errors in parentheses: p b 0.01, p b 0.05, and p b 0.1.

change.30 This effect is measured by the interaction between the share


of individuals aged 60 and older over the total population in department
i at the start of the period [t 1, t] and the share of individuals
employed in routine positions in department i at the start of the period
[t 1, t]. The value of both shares is taken at the start of the period to
capture the exogenous effect of aging on the growth of the personal service employment share in the presence of technological change. If our
hypothesis is conrmed by the data, we expect the coefcient of the
interacted variable to be positive.
Our benchmark equation stacks the two time periods discussed earlier, and includes a full set of time effects, department effects and a
proxy of their interacted effect. Department dummies allow us to capture time-invariant determinants of the share of service employment
across departments, such as geographical characteristics (e.g. weather,
sea, proportion of urban and rural areas). Time dummies allow us to
capture time-varying determinants of the share of service employment
common to all departments (this mainly concerns institutional
reforms).
Because our regressions exploit heterogeneity across French departments, the impact of national institutional reforms should be captured
by time dummies. However, we could argue that a given institutional
change may have heterogeneous effects across departments. Ideally,
we would introduce time and department interacted xed effects to
control for this time-varying spatial heterogeneity. However, this
would raise the number of control variables above the number of observations. We thus proxy these interacted xed effects by the proportion
of individuals earning the minimum wage (between 1 and 1.1 times the
minimum wage) in each department. This variable changes from one
year to the next and from one department to another. Therefore, it allows us to control for the time-varying heterogeneous effect of national
reforms across departments. Furthermore, individuals earning the minimum wage are likely to have been affected by major institutional reforms in France during the period considered. One such reform was

30
In Autor and Dorn (2013), the authors introduce the share of seniors in a commuting
zone as a control variable of a regression trying to estimate the impact of the routine employment share in the growth of service employment. They recognize that they are unable
to provide an explanation for the negative and signicant coefcient associated with the
share of seniors, since they expected this coefcient to be positive.

the minimum wage itself, which evolved signicantly around 2000. Another reform was the introduction of the RSA31 in 20072008. The
beginning-of-period values of four additional explanatory variables32
(denoted by Xi,t 1) are also included in our specication:
1. The share of managers33 in department i. This variable seeks to capture the income effect, invoked by Mazzolari and Ragusa (2013) and
Manning (2004), responsible for the increase in demand for lowskilled personal services. This variable also controls for another effect
highlighted in Manning (2004), according to whom the employment
of low-skilled workers is increasingly dependent on physical proximity to high-skilled workers.
2. The unemployment rate measures local labor market conditions. The
expected sign of this coefcient is unknown a priori as a high unemployment rate can have divergent effects. On the one hand, when the
unemployment rate is high the average level of income that can be
spent on service activities is low, thereby decreasing demand for personal services (such as gardeners, house cleaners, child care). On the
other hand, a high unemployment rate may promote a more intensive labor switch towards the service sector. Depending on which effect dominates, the coefcient will be positive or negative.
3. The share of manufacturing jobs is another indicator of local labor
market conditions. To avoid collinearity problems, from the set of
manufacturing jobs we discard those that are already included as
routine in the regression. So, in the share of manufacturing jobs, we
actually nd only the share of non-routine jobs in the manufacturing
sector. The expected sign of this variable is ambiguous. On the one
hand, it controls for a composition effect. The service sector is likely
to expand more slowly if there is a large share of non-routine
31
Revenu de Solidarit Active. Minimum income guaranteed to all French individuals fullling a given number of characteristics.
32
By taking the value at the beginning of the 9-year period, we essentially avoid
endogeneity issues. Table 6 in the web Appendix displays the estimation results obtained
when the explanatory variables are dened in variation and shows the problems linked to
this approach.
33
The INSEE (National Statistical Institute) classication includes the following list of
managers: engineers, managers from the public and private sector, professors, scientic
occupations, administrative managers, marketing managers, journalists, and jobs in the
arts and life performance.

E. Moreno-Galbis, T. Sopraseuth / Labour Economics 27 (2014) 4455

manufacturing employment. On the other hand, it may also capture


an income effect since people employed in non-routine manufacturing positions may ask for personal services.
4. The share of the active female labor force34 constitutes another potential demand shifter. Many services, such as restaurant meals or
housekeeping, serve as substitutes for household production.
Hence, higher female labor force participation might be expected to
raise demand for these services (see Manning (2004) or Mazzolari
and Ragusa (2013)). On the other hand, many personal service jobs
are implemented by women so lower female labor force participation may prevent them from providing some personal services.
All variables are computed using the French LFS except the share of
old workers and the unemployment rate, which are taken from the
French Institute of Statistics database (INSEE).35
Controlling for xed effects (i.e. department, year and a proxy of the
interacted xed effect), columns 1 and 2 of Table 2 report the estimates
on the individual impact of the proportion of seniors and the share of
routine jobs on the demand for services. While the share of seniors on
its own displays a non-signicant positive effect, the coefcient associated with the routine variable is positive and signicant. Population aging
alone does not seem to have a signicant effect on the demand for personal services, while, consistent with the literature, the routinization
hypothesis displays a small but positive and signicant coefcient.
Column 3 interacts both population aging and the routinization
hypothesis. The new variable better matches Eq. (9) which accounts
for the divergent effect of technological progress resulting from relationship between goods and personal services (i.e. whether they are
compliments or substitutes). The estimated coefcient on this variable
is positive, signicant and displays a larger value than the coefcient associated with the individual variables oldt 1 and routinet 1.
In column 4, we add to the benchmark regression an alternative demand shifter: the share of managers in the department. By controlling
for managers, we aim to capture the income inequality effect outlined
in Mazzolari and Ragusa (2013) and Manning (2004). The coefcient
on the main explanatory variable remains essentially unaffected by
the introduction of the share of managers. The estimated coefcient associated with this variable is positive but not signicant. The increase in
employment share in the personal service sector does not seem to respond to the income inequality effect resulting from the increased presence of highly-paid managers in the department (this is consistent with
Autor and Dorn (2013) and Goos et al. (2009)). Rather, it responds to
the demand shift explained by the combined effect of technological
change and aging.
Column 5 introduces two additional variables to control for local
labor market conditions: the unemployment rate and the share of
non-routine manufacturing employment in the department. Whereas
the former is not signicant, the proportion of non-routine manufacturing employment displays a positive and signicant effect, which could
be interpreted as an income effect.
Finally, column 6 controls for female labor force participation. This
variable does not display a signicant role and its inclusion does not
modify the size or signicance of the rest of the coefcients.
Additional robustness tests are included in a web Appendix
(see http://thepthida.sopraseuth.free.fr, https://sites.google.com/site/
morenogalbiseva). We report estimates for alternative regression specications or denitions of routine occupations. We also control for the
potential effect associated with the approval in July 2005 of the Law

34
A woman is considered to belong to the active population if she is employed or actively searching for a job (unemployed). The share of the active female labor force is thus dened as the number of active women divided by the female working-age population.
35
In the web appendix we report the descriptive statistics on the variables used for the
regressions (see http://thepthida.sopraseuth.free.fr, https://sites.google.com/site/
morenogalbiseva).

53

for the development of Personal Services. Estimations from Table 2,


as well as the robustness tests provided in the web Appendix, conrm
that, in a context of technological progress, population aging has promoted employment growth in the personal service sector. This effect remains signicant after controlling for the potential impact of other
demand shifters such as the share of highly paid managers, the share
of workers in non-routine manufacture tasks, and institutional changes.
6. Conclusion
Over the past years, the economic literature has faced difculties in
explaining the reasons behind the observed polarization of the labor market between lovely and lousy jobs. While a general consensus exists
around the idea that biased technological progress is the main cause behind the progressive concentration of jobs at the top of the wage distribution, the concentration of jobs at the bottom of the wage distribution is
more difcult to explain. Contrary to jobs at the top of the distribution,
productivity of positions at the bottom of the distribution has not been
improved by technological progress, so the rise in demand for this type
of job is more difcult to understand. A recent explanation proposed by
Autor and Dorn (2013) states that if goods and personal services are complements, the diffusion of new technologies yields a rise in demand for
personal services since the relative price of goods falls.
Autor and Dorn (2013) do not to provide an explanation as to why
we do not observe any polarization of the labor market before the
mid-1990s, even though American technological diffusion started in
the 1980s. Our paper points to population aging as the main reason behind the increased complementary relationship between goods and
personal services. Using the French Household Budget Survey, we estimate the elasticity of substitution between goods and personal services
for young and old households. Our estimations reveal that goods and
personal services tend to be complementary for seniors and substitutable for young households. Then, if the elasticity of substitution is constant, population aging is enough to explain the reduction in the
elasticity of substitution at the aggregate level.
This increased complementary relation between goods and services
since the mid-1990s was associated with a reduction in the relative
price of goods resulting from the automation of routine jobs. This caused
an increase in demand for personal services. Using the share of seniors
as a proxy for the aggregate complementarysubstitutability relationship between goods and personal services, we estimated the combined
impact of population aging and technological progress on the demand
for personal service employment. The estimations conrm our initial
guess: the combined effect of population aging and technological diffusion has increased demand for personal services and, thus, demand for
labor in these positions.
Appendix A. Aggregate elasticity of substitution between goods and
personal services
The aggregate demand for personal services is:
Y

qs M Y qs ps M O qs ps

14

where Y and O refer to Young and Old respectively. MX is the number of


persons X and qXs (ps) the demand from a person of population X. From
Eq. (14), we get
"
#
dqs ps ps
dqY p
dqO p

MY s s MO s s
dps
dps
dps qs qs
or
aggregate

M Y qYs young M O qOs old

qs
qs

with

dqXs ps
:
dps qXs

54

E. Moreno-Galbis, T. Sopraseuth / Labour Economics 27 (2014) 4455

Table 3
Elasticity of substitution between goods and personal services (all expenditures on
personal services). Budget Household Survey 2006.

Log y
Log ps
Geographical dummies
Observations
R-squared
VIF

Table 4
Elasticity of substitution between goods and personal services (house cleaning). Pooled
sample. Budget Household Survey 2006.

Dependent variable: demand for all personal services

Dependent variable: Demand for house cleaning services

All senior households

All young households

(1)

(2)

(3)

(4)

All senior and young


households

All senior and


young households

0.124***
(0.000929)
0.565***
(0.00170)
Yes
2370
0.041
16.63

0.133***
(0.000927)
0.574***
(0.00171)
No
2370
0.018
1.01

0.337***
(0.000848)
1.081***
(0.00236)
Yes
1801
0.109
13.77

0.332***
(0.000834)
1.095***
(0.00237)
No
1801
0.092
1.00

Standard errors in parentheses: p b 0.01.

The aggregate elasticity of substitution between goods and personal


services aggregate is a weighted average of the elasticities of substitution
of young (young) and old (old) consumers. The weight of old is
MO qOs
M O qOs
1

M q
:
qs
MY qYs M O qOs
M q 1
Y

Y
s
O
s

15

If, in a geographical area with a large fraction of old workers (M0 is


large compared to MY), the weight in front of old in Eq. (15) tends to
be 1, then the aggregate elasticity of substitution between goods and
personal services tends to be that of senior consumers old.
Appendix B. Robustness tests on the estimation of the elasticity
B.1. All expenditures on personal services
For each household, we aggregate here all annual expenditures on
domestic services (house cleaning, child care, shopping, domestic
health care, gardening and other tasks) declared by the household as
well as the aggregate number of hours worked by the personal service
employees. Due to the low number of households declaring strictly positive expenditures, we are obliged to consider a composite of services
and we are unable to estimate the elasticities of substitution individually for each type of personal service. To compute the hourly price paid by
households for personal services, we follow the same procedure as for
the cleaning services.36 Results are reported in Table 3.
B.2. Pooled sample
In Table 1, even if our estimations reveal that the elasticity of substitution between goods and personal services is always larger for young
households than for senior households, in some cases, the coefcients
remain quite close. In order to test if the estimated elasticities are significantly different for young and senior households, we consider a pooled
sample including all of them. First, we add an interacted term resulting
from multiplying the hourly price of services and a dummy taking the
unit value when considering a senior household (columns 1 and 2 of
Table 4). Secondly, we introduce an interacted term resulting from multiplying the income and the dummy variable of seniors (columns 3 and
4 of Table 4). We implement both regressions in the presence and in the
36
First, since we deal with yearly expenditures, in order to compute the hourly cost we
are obliged to assume that the number of hours worked by the personal service employee
in one year is six times the number of hours worked in two months. Secondly, to deal with
the problem of missing data in the number of hours, we compute the average number of
hours hired by different categories of household classied according to their yearly
expenditure. This average number of hours is then imputed to the households of the corresponding spending category if the number of hours is missing. Finally, to compute the
hourly price paid by households for personal services, we divide yearly expenditure by
yearly hours (declared or imputed). We drop observations associated with extreme
observations.

Log y
Log ps
Log ps senior

(1)

(2)

(3)

(4)

0.196***
(0.000632)
1.263***
(0.00139)
0.349***
(0.000556)

0.201***
(0.000628)
1.261***
(0.00139)
0.350***
(0.000556)

Yes
4164
0.092
15.01

No
4164
0.081
1.27

0.217***
(0.000725)
1.412***
(0.00297)
0.528***
(0.00321)
0.0353***
(0.000623)
Yes
4164
0.093
17.06

0.222***
(0.000721)
1.417***
(0.00298)
0.538***
(0.00321)
0.0371***
(0.000624)
No
4164
0.081
22.93

Log y senior
Geograph. dummies
Observations
R-squared
VIF

Standard errors in parentheses: p b 0.01.

absence of geographical dummies. The value and signicance of the estimated coefcients remains fairly similar, suggesting that collinearity is
not a major problem.
In columns 1 and 2, the interacted term log ps senior arises as positive and signicant. To determine the particular elasticity associated
with seniors, the coefcient of the individual variable must be added
to that of the interacted variable. We nd that the coefcient associated
with log ps is signicant above one37 (in absolute terms) at a signicance level of 1%. When adding this coefcient to that of the interacted
variable, we obtain a value below one, conrming the idea that goods
and personal services are complements for seniors and must then be
substitutes for young households, since the average estimated elasticity
over the pooled sample is above 1.
In columns 3 and 4, we introduce two interacted terms log ps
senior and log y senior. Again, the coefcient associated with ps is
signicant above one.38 As in columns 1 and 2, when added to the
coefcient associated with log ps senior the obtained value is below
1 implying the complementary relation between goods and personal
services for seniors. The coefcient associated with the income variable
arises as positive and signicant. However, when interacted with the
senior dummy, the interacted term displays a negative and signicant
coefcient, suggesting that, for seniors, the demand for services is relatively less dependent on the level of income with respect to the pooled
sample (probably because for many seniors there is a need for personal
services). In spite of the stability of the value and signicance of the
coefcients between columns 3 and 4, note that the introduction of
the interacted term log y senior induces a multicollinearity problem
that does not disappear even when we eliminate the geographical
variables.39
Appendix C. Denition of jobs
Farmers, civil servants, the military and clergymen are excluded. All
jobs related to these categories are dropped from the sample.
Some occupations are characterized by high employment and a
vague denition. Following Goos and Manning (2007), each of the following jobs is dened as the specic occupation in a particular industry
(15 sectoral activities): secretary; clerks in nancial departments and
37
Actually 1 does not belong to the condence interval associated with 1.263 or
1.261 which conrms that the coefcient is signicantly above 1 in absolute terms.
38
Again, 1 does not belong to the condence interval associated with 1.412 or
1.417.
39
Since the samples are different, and constant terms differ between estimations in columns 34 of Table 4 and estimations in Table 1, the estimated coefcients from Table 4
will necessarily be different from those in Table 1.

E. Moreno-Galbis, T. Sopraseuth / Labour Economics 27 (2014) 4455


Table 5
Change in log employment (dlnN) over 19932010 for occupations ranked by their
median 1993 wage (lnW1993).

(a) Top paying jobs


Flying personal
Managers in big rms, administrative, nancial
or sale department
Technical managers in big rms
Engineer and executive managers

lnW1993

dlnN

10.13204
10.006

.0465518
.006904

9.954719
9.87817

.0624228
.2712399

8.834143

.0116572

8.836745

.0167217

(b) Middling occupations


Ofce clerk in accounting and nancial department,
transportation industry
Ofce clerk in accounting and nancial department,
food industry
Ofce clerks, misc, intermediate good industry
Skilled worker in weather-stripping
Skilled worker in extraction
Secretaries, engineering industry
Unskilled worker in sorting, wrapping and delivery
activities, engineering industry
Secretaries, car industry
Secretaries, food industry

8.838068
8.839276
8.839276
8.839276
8.844096

.00573
.0208104
.1188045
.0439425
.0244403

8.847156
8.853665

.0353632
.0379203

(c) Lowest paying jobs


Cleaners, nancial industry
Transportation services in health industry
Cleaners, consumption goods industry
Housework employees in private homes
Child care

7.699993
7.740664
7.775527
7.783224
7.824046

.1208007
.0678637
.0386493
.0086223
.0827386

(d)
Top 10%
Middling occupation (4555%)
Bottom 10%
Service occupations

9.618954
8.827257
8.074184
8.30726

.0466335
.0040566
.0277598
.0376211

accounting; cleaners; clerks in various departments; unskilled workers


in mechanical works, monitoring and assembly line; courier, messenger; warehouseman, unskilled workers in shipping and transportation;
unskilled worker in sorting, wrapping and delivery activities; drivers;
executive workers in small rms in nance or administration;
storekeeper.
The 15 sectors are: Food industry; Consumption Goods; Car industry; Engineering industry; Intermediate goods; Energy; Construction;
Trade and repairing; Transportation; Financial activities; Real estate;
Services to rms; Services to private individuals; Education, health,
social services; and Administration.
Some jobs may have disappeared, while new ones are emerging. The
French LFS modied the job classication in 2003 in order to take into
account the changes in occupations. We paid attention to having a consistent denition of jobs throughout the 18 years of our sample. There
are no new occupations that cannot be included in the pre-2003
classication.

Appendix D. Descriptive statistics on job polarization


Table 5 ranks occupations from highest-paid to lowest-paid. For
each occupation, Table 5 presents the average log change in employment and wages. Among the fastest expanding occupations (panel (a)
of Table 5) are managers and executives but also the lowest-paid jobs
(panel (c)) such as child care and transportation services. The biggest

55

declines in employment (panel (b)) are observed for middling occupations such as secretaries, ofce clerks and some skilled workers in specic industries. Panel (d) of Table 5 presents the average change in log
employment over 19932010 for the top 10% jobs (ranked by the median 1993 wage), bottom 10% and middling jobs (identied as jobs in the
4555% percentiles of the 1993 wage distribution). The table conrms
the fast employment growth at the top and bottom of the job quality
ladder while middling occupations have been characterized by a decline
in employment. Furthermore, most of the employment growth observed at the bottom of the wage distribution comes from service occupation jobs.
Appendix E. Data for regressions
In the sample survey, 96 departments are considered. In all regressions, departments are weighted by their total population. Jobs included
in service occupations are: waiter/ waitress; hotel staff; receptionist; accompanying services; janitor; cleaner in services for private individuals;
gardener; cleaner and staff in private homes; child care, family worker;
auxiliary nurse; workers in hospitals; ambulance driver; manicure,
beautician (employee); hairdresser (employee); manicure, beautician,
hairdresser (self-employed); worker in various services to private
individuals.
For want of an accurate description of the tasks involved in each job,
we choose among the list of 452 occupations a set of jobs considered as
routine. In order to establish this list, we look at routine jobs identied
in Autor and Dorn (2013) and Spitz-Oener (2006). The following occupations are considered as routine jobs: typist, keyboarder; operator on
computer; ofce clerk in accounting and nancial department in all industries; ofce clerk in misc. departments in all industries; unskilled
workers in mechanics; unskilled workers in extraction; unskilled
workers in wrapping, delivery. We check that these jobs are characterized by a 1993 median range that lies in the middle of the 1993 wage
distribution (4060% range).
References
Autor, D., Dorn, D., 2013. The growth of low-skill service jobs and the polarization of the
U.S. labor market. Am. Econ. Rev. 103 (5), 15531597.
Autor, D., Levy, F., Kearney, M., 2006. The polarization of the US labor market. Am. Econ.
Rev. 96 (2), 189194 (January).
Autor, D., Katz, L., Kearney, M., 2008. Trends in the U.S. wage inequality: re-assessing the
revisionists. Rev. Econ. Stat. 90 (May), 300323.
Baumol, W., 1967. Macroeconomics and unbalanced growth: the anatomy of urban crisis.
Am. Econ. Rev. 62, 415426.
Cheron, A., Langot, F., Moreno-Galbis, E., 2011. Labor market institutions and technological employment. Economica 78 (309), 159186.
Clark, C., 1957. The Conditions of Economic Progress. Macmillan, London.
Goos, M., Manning, A., 2007. Lousy and lovely jobs: the rising polarization of work in
Britain. Rev. Econ. Stat. 89 (1), 118133 (February).
Goos, M., Manning, A., Salomons, A., 2009. Explaining job polarization in Europe: the roles
of technology and globalization. Am. Econ. Rev. Pap. Proc. 99 (22), 5863.
Hirsch, B.T., Berger, M.C., 1984. Union membership determination and industry characteristics. South. Econ. J. 50 (3) (January).
Landais, C. 2008. Top income in France: booming inequalities. Paris School of Economics,
Mimeo.
Lemieux, T., 2006. Increasing residual wage inequality: composition effects, noisy data or
rising demand for skill? Am. Econ. Rev. 96 (3), 461498.
Manning, A., 2004. We can work it out: the impact of technological change on the
demand for low-skill workers. Scott. J. Polit. Econ. 51, 581608.
Maurin, E., Thesmar, D., 2004. Changes in the functional structure and demand for skills.
J. Labor Econ. 22 (3), 639664.
Mazzolari, F., Ragusa, G., 2013. Spillovers from high-skill consumption to low-skill labor
markets. Rev. Econ. Stat. 95 (1), 7486 (March).
Spitz-Oener, A., 2006. Technical change, job tasks and rising educational demands:
looking outside the wage structure. J. Labor Econ. 24 (2), 235270.